Beyond One-hot Encoding: lower dimensional target embedding
This addresses the issue of inefficient training in convolutional neural networks for large-scale classification tasks, offering a method to exploit label relationships, though it is incremental as it builds on existing encoding strategies.
The paper tackles the problem of one-hot encoding ignoring label relationships in large-scale datasets by embedding targets into a low-dimensional space, improving convergence speed while preserving accuracy, with experiments showing drastic convergence improvements and competitive accuracy on datasets like CIFAR-100 and ImageNet.
Target encoding plays a central role when learning Convolutional Neural Networks. In this realm, One-hot encoding is the most prevalent strategy due to its simplicity. However, this so widespread encoding schema assumes a flat label space, thus ignoring rich relationships existing among labels that can be exploited during training. In large-scale datasets, data does not span the full label space, but instead lies in a low-dimensional output manifold. Following this observation, we embed the targets into a low-dimensional space, drastically improving convergence speed while preserving accuracy. Our contribution is two fold: (i) We show that random projections of the label space are a valid tool to find such lower dimensional embeddings, boosting dramatically convergence rates at zero computational cost; and (ii) we propose a normalized eigenrepresentation of the class manifold that encodes the targets with minimal information loss, improving the accuracy of random projections encoding while enjoying the same convergence rates. Experiments on CIFAR-100, CUB200-2011, Imagenet, and MIT Places demonstrate that the proposed approach drastically improves convergence speed while reaching very competitive accuracy rates.